Hopefully, the number of genes with cross-mapping activity is negligible.
If you want to be conservative, you can ignore all genes that exhibit *any* cross-mapping.
However, I suspect this won't make much of a difference, if the number of genes removed is small.
You'll still get cross-mapping elsewhere due to sequencing errors, exon junction reads, etc. which will probably make up the bulk of errors.
We're also probably missing stuff because we haven't looked at every base in the genome, only extracting read sequences every 5-10 bp.

I guess the conclusion here is that the vast majority of transcriptomic or genomic regions do not have cross-mapping.
The actual effect of cross-mapping depends on the level to which the offending regions are transcibed, which is unknown.
In practice, the requirement for unique mapping should provide minimize cross-mapping by refusing to align reads with two candidate locations on different genomes.
You'd only really have a threat if you get a mutation that makes it more like another species, which seems unlikely.

For the purposes of the spike-in experiments, you might get extra variability due to cross-mapping from variable mouse genes to either spike-in set.
However, this should be observable in the control experiment, so you should be able to subtract it out fairly easily.
Cross-mapping between spike-in populations, or from the mouse to both spike-in populations, is more troublesome; variances would shrink to zero.
Best results may be obtained by keeping the ERCCs as one population in all experiments, as they're pretty robust to cross-mapping.